We tested phase message transport in seven public SEEG rest records from OpenNeuro ds006065 version 1.0.0. A frozen local reader used 100 ms of past voltage to forecast phase 50 ms ahead. Each replay had 11,600 outputs and four pairs. We erased exactly 10% of time slots as isolated 10 ms losses or 50 ms bursts. Each person had 100 loss schedules. A sparse rule sent a phase when local error exceeded 10 degrees. Reliable acknowledgments before the next output restored the 10 degree local agreement bound at every available output, under immediate forward delivery and a shared clock. Full updates also met this bound and needed no return traffic. Under bursts, acknowledgments reduced the all-output fraction above 10 degrees from 8.97% to 7.87%. The paired change was -1.097 percentage points, with a descriptive 95% person interval of -1.220 to -0.965. The six-byte acknowledgment rule used 16.27% more total bytes than full updates; its interval was 14.38% to 18.19%. A fresh full-source audit and an independent scalar and wire replay passed. The package contains voltage, models, loss masks, scores, actual messages, figures, proof, code, and reproduction steps. The cohort was reused. Loss was simulated. The local reader had 56.56 degrees of mean error to an offline reference before loss. The bound concerns agreement with that reader, not true neural phase. No physical link or stimulation was tested. This version corrects the checksum tool after archive download.
Filters: arxiv: cs.NI (Networking and Internet Architecture) ×
We tested a method to send a continuous brain reader score with a stated error limit. The source was public OpenNeuro ds004473 version 1.0.2. We used 67,309 later SEEG windows from eight people. The sender sent four-bit power codes. The receiver requested selected original values once. It released a score only when an outward interval proved its error limit. The main limit was 5 percent of the score standard deviation in early fitting data. Every later window passed. Mean cost was 108.1 bytes, compared with 128 bytes for exact feature transfer. The mean saving across people was 15.5 percent. The descriptive 95 percent person interval was 10.3 to 21.2 percent. At 16 extra bytes per message, the method cost 8.4 percent more than exact transfer. A 1 percent score limit also cost more before message overhead. The earlier binary decision rule met the 5 percent score interval limit in only 9.6 percent of windows on average. The method assumes fixed affine readers, finite float32 features, shared parameters, and reliable ordered replies. The proof gives an error bound for each released score. It does not require independent SEEG windows. The package contains the proof, source recipes, fixed data, all masks and bounds, three figures, code, and checks. Independent wire replay passed 201,927 score checks. Fresh source checks covered all eight people and 28 raw windows. This study adds a score transport method. It does not establish better clinical decoding, stimulation, or radio power.